[FFmpeg-devel] [PATCH] start code detection for hardware decode in mpeg4/divx streams

[anuj mittal]

On Thu, Oct 18, 2012 at 11:45 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
> On Thu, Oct 18, 2012 at 09:24:11PM +0530, anuj mittal wrote:
>> On Wed, Oct 17, 2012 at 4:26 PM, anuj mittal <am.devel at gmail.com> wrote:
>> > Hello
>> >
>> > I've some AVIs with divx bitstream packed frames. When I play these
>> > file using hardware decode VAAPI, I notice frame drops.
>> >
>> > It looks like that the logic to calculate the end of the frame (while
>> > calculating the slice size to be sent to hardware) isn't correct for
>> > hardware decode. If I have a B frame packed with a P, it just sends
>> > the size of P+B to the hardware. Because of this, the packed B frames
>> > do not get decoded at all.
>> >
>> > The attached patch tries to fix this behaviour. The code will now
>> > check for the start code while calculating the value of "end". If the
>> > start of next frame has been detected, it will update s->mb_y.
>> >
>> >
>>
>> Does this look ok? Thank you.
>>
>> Anuj
>
>>  h263dec.c    |   11 ++++++++++-
>>  ituh263dec.c |   14 ++++++++++++--
>>  2 files changed, 22 insertions(+), 3 deletions(-)
>> b098443d3108d1d133b562088791aaa8bbdf6826  0001-Start-code-frame-handling-while-decoding-using-hardw.patch
>> From c654d9a2aca9f5f20c2949f18ce584eb363c47e5 Mon Sep 17 00:00:00 2001
>> From: Anuj Mittal <am.devel at gmail.com>
>> Date: Wed, 17 Oct 2012 16:08:42 +0530
>> Subject: [PATCH] Start code frame handling while decoding using hardware.
>>
>> ---
>>  libavcodec/h263dec.c    |   11 ++++++++++-
>>  libavcodec/ituh263dec.c |   14 ++++++++++++--
>>  2 files changed, 22 insertions(+), 3 deletions(-)
>>
>> diff --git a/libavcodec/h263dec.c b/libavcodec/h263dec.c
>> index 7141d49..4602f8d 100644
>> --- a/libavcodec/h263dec.c
>> +++ b/libavcodec/h263dec.c
>> @@ -169,10 +169,19 @@ static int decode_slice(MpegEncContext *s){
>>      ff_set_qscale(s, s->qscale);
>>
>>      if (s->avctx->hwaccel) {
>> +        int ret;
>> +
>>          const uint8_t *start= s->gb.buffer + get_bits_count(&s->gb)/8;
>>          const uint8_t *end  = ff_h263_find_resync_marker(s, start + 1, s->gb.buffer_end);
>>          skip_bits_long(&s->gb, 8*(end - start));
>> -        return s->avctx->hwaccel->decode_slice(s->avctx, start, end - start);
>> +        ret = s->avctx->hwaccel->decode_slice(s->avctx, start, end - start);
>> +
>> +        /* Next frame detected */
>> +        if (!end[0] && !end[1] && (end[2] == 1)){
>
> these can access out of array values

Thanks for pointing this out. I am now checking for (end <=
s->gb.buffer_end-2). It should be ok now I think.

>
>> +            s->mb_y = s->mb_height;
>> +        }
>> +
>> +        return ret;
>>      }
>>
>>      if(s->partitioned_frame){
>> diff --git a/libavcodec/ituh263dec.c b/libavcodec/ituh263dec.c
>> index 6d43233..f4470dc 100644
>> --- a/libavcodec/ituh263dec.c
>> +++ b/libavcodec/ituh263dec.c
>> @@ -217,11 +217,21 @@ const uint8_t *ff_h263_find_resync_marker(MpegEncContext *s, const uint8_t *av_r
>>          int prefix_len = ff_mpeg4_get_video_packet_prefix_length(s);
>>          for(;p<end; p+=2){
>>              if(!*p){
>> -                if      (!p[-1] && ((p[1] >> (23-prefix_len)) == 1)) return p - 1;
>> -                else if (!p[ 1] && ((p[2] >> (23-prefix_len)) == 1)) return p;
>> +                if      (!p[-1] && (((p[1] >> (23-prefix_len)) == 1) || (p[1] == 1))) return p - 1;
>> +                else if (!p[ 1] && (((p[2] >> (23-prefix_len)) == 1) || (p[2] == 1))) return p;
>
> this isnt a resync marker, so it does something else than what the
> name of the function would suggest
>

I've changed this to a more generic name. Attached is the patch with
changes. Thank you.

Anuj
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Start-code-frame-handling-while-decoding-using-hardw.patch
Type: application/octet-stream
Size: 3713 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20121019/488e1355/attachment.obj>